Answers are grounded in this paper and cite their sources.
Time-critical autonomous and adaptive systems increasingly deploy and execute security-sensitive decision gates directly on edge devices, where symmetric encryption algorithms are applied to protect critical data or achieve secure communication. However, implementations of symmetric cryptographic algorithms typically rely on lookup-centric white-box cryptography (WBC) components.In hostile Man-At-The-End (MATE) deployment environments, the discrete table structures of such components can expose regular intermediate representations to automated reverse engineering and leakage-model attacks such as Differential Computation Analysis (DCA). This motivates the investigation of alternative implementations that reduce alignment with table-oriented analysis while preserving exact functional behavior. We present NeuroMask, a time-critical hardening neural implementation of secret-dependent byte lookups that replaces table access with a compact continuous surrogate and restores exact discrete semantics via a deterministic integer snapping rule. This shifts protected computation from boolean/array logic to mixed floating-point and transcendental operations while keeping a predictable runtime profile. We instantiate NeuroMask with SIREN-based surrogates trained using full-batch L-BFGS, achieving bit-perfect reconstruction of standardized AES and SM4 S-Boxes. Our evaluation combines correlation-based DCA under a controlled leakage model with latency and throughput measurements, positioning NeuroMask as a practical pattern for certified, security-hardening components in autonomous edge applications.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: